I think a key thing here is "plan to systematically do something that doesn't scale if the rewards are big enough". And the common bias is thinking that the rewards are not as big as they are. If it's in your top 3 problems and manual work is a path to scaling automatically, you should probably do it.

However, at some point short-term velocity is also contributing speed bumps that are barriers for reaching the next level. I certainly wouldn't want to be doing this ad-hoc or without consideration.